This program is intended to support doctoral students who are advanced to candidacy at the time of nomination by their department to the Division of Graduate Education. Approximately 160 fellowships are awarded under this program yearly across the UCLA campus. Atefeh's dissertation is on customized computation and machine learning. Domain-specific accelerators (DSA) can offer high performance while being energy efficient since the designer can customize all the hardware parameters such as data types, memory access, parallelism, control/data path, etc. However, they are not easy to generate, as the designer must describe their architecture at the circuit level. This has limited the DSA community to hardware designers and prevented their large adoption. We aim to alleviate this problem by combining customized computing and ML. More specifically, our efforts consist of two parts: 1) customized computing for ML applications. Because of their wide usage, we develop architecture templates for accelerating them to decrease their development cycle. 2) ML techniques to automate the optimization of customized computing for general applications. We formulate their design space and develop dedicated heuristics to efficiently explore them. Moreover, we develop ML models to predict the quality of designs in milliseconds rather than minutes/hours taken by the HLS tools. This research can open new doors to those without hardware knowledge to exploit DSAs which in turn helps to broaden its community and further improve its technology.
